rockycode

Visualize a neural-net block (attention, transformer, MoE, diffusion-LLM) as an artifact three linked ways — the paper diagram, the tensor-shape ribbon, and the user's PyTorch — cross-highlighted so hovering one lights the other two. Grounded in the real paper/code, never drawn from memory.

Formalize and machine-verify mathematics AND neural networks in Lean 4 — Mathlib proofs for the math, TorchLean for typed models and robustness certificates, honest green/amber/red verdicts, rendered as an artifact